Research On Key Issues Of Pedestrian Detection In Road Scenarios

Pedestrian detection is one of the research hotspots in the field of computer vision.It is the basis of tasks such as autonomous driving,human-computer interaction,and behavior recognition.It has important research value and a wide range of application scenarios.The vehicle-mounted pedestrian detection algorithm automatically detects the position of the pedes-trian in front based on the image data to remind the driver or the automatic driving system to avoid collision events.With the rapid development of computer processing capabilities,deep learning theories and technologies have achieved remarkable results and demonstrated unlim-ited potential in many fields,and the field of pedestrian detection has also made great progress.However,due to the challenges of real road scenes,changes in lighting,and diversity of occlu-sion,reducing false alarms and missed detections is always a problem for pedestrian detection.Carrying out special research based on deep learning to improve pedestrian detection perfor-mance and reduce missed detections and false alarms is a top priority for autonomous driving and road traffic intelligence.This paper focuses on the research on the challenging problems of pedestrian detection in road scenes.Far-infrared vehicle pedestrian detection is an effective way to avoid collisions between people and vehicles at night.However,due to the lack of large-scale far-infrared refer-ence data sets,it is difficult to effectively train detection models,which restricts the development of this field.Therefore,this paper proposes,establishes and discloses a large-scale far-infrared pedestrian detection benchmark data set.False alarm suppression is a difficult problem of ve-hicle pedestrian detection in road scenes.The existing framework based on the strategy of full-image sliding windows leads to a large number of pedestrian false alarms that deviate from the road surface area.This paper studies and proposes a vehicle pedestrian detection network based on road surface semantics.Pedestrian occlusion is an important challenge in the field of pedestrian detection in complex road scenes.In this paper,pedestrian detection methods based on human visible information are studied,and a method based on head clues to gradually expand the perception domain is proposed to improve the recall rate of severely occlusion pedestrians.The main contributions of the paper are as follows:1)For vehicle night pedestrian detection,a large-scale far-infrared vehicle pedestrian de-tection data set is proposed and established.With the basic goal of maintaining authenticity,diversity and challenge,with the help of vehicle-mounted far-infrared sensors to collect video data of various road conditions,formulate specifications to manually annotate the data,includ-ing multiple scenes,multiple road sections,a total of 210,000 frames of images and 470,000 Pedestrians mark boxes,so the data scale can meet the data volume requirements of deep neural network training and evaluation tests.Statistical analysis was performed on the annotated data to obtain characteristics such as pedestrian scale distribution,aspect ratio distribution,distance distribution,and center distribution,which provided basic prior knowledge for the design of vehicle pedestrian detection algorithms.Modify the key parameters of the classical pedestrian detection neural network framework,design and complete detailed ablation experiments,and design a benchmark detector for this data set.Experimental results show that the benchmark detector is superior to the existing advanced pedestrian detection algorithm on the data set es-tablished in this paper.2)Aiming at the false alarm problem in pedestrian detection,based on the characteristics of pedestrian distribution in driving scenarios,a pedestrian detection network based on road surface semantics is proposed.Based on a two-stage pedestrian detection framework,by designing a lightweight road surface area estimation branch,road surface horizontal center area estimation based on shared convolution features is realized.Based on the template convolution method,an area recommendation network based on the horizontal center area of the road surface is realized.Experimental results show that the network proposed in this paper can accurately estimate the pedestrian center distribution area,and recommend pedestrian candidate areas based on the pavement area can significantly reduce the false alarm and reduce the computational cost.On the day and night data and multi-spectral data,the methods proposed in this paper have achieved a reduction in the missed detection rate,proving that the integrated road surface semantic pedestrian detection framework proposed in this paper can effectively deal with the problem of deviation from the road false alarm.3)Aiming at the problem of pedestrian occlusion in complex road scenes,a human cas-cade pedestrian detection network based on pedestrian head and visible information features is proposed.The occlusion problems in complex road scenes are divided into pedestrian oc-clusion problems and inter-occlusion problems caused by background objects.Aiming at the problem of intra-class occlusion,a center point estimation and scale regression area recommen-dation network without anchor point matching is proposed to achieve feature response sharpen-ing and reduce feature adhesion? separately establish pedestrian head and pedestrian center area recommendation modules to realize pedestrian interest areas(Region of Interests,Ro Is)com-plement each other to improve the occlusion pedestrian recall rate? to address the problem of inter-occluded objects interfering with Ro Is pooling features,a visible area guidance attention module is proposed.By estimating the probability distribution map of pedestrian visible areas,the Ro Is pooling features After re-weighting,the noise characteristics of occlusion objects are suppressed? for the problem of lack of distinguishable information in the head Ro Is,a head cas-cade detection network is proposed,and the perceptual domain resampling strategy based on the cascade detection module is gradually expanded in stages to achieve head Partial clues re-duce the serious omission of pedestrians,while avoiding the problem of introducing more false alarms due to the small head size and lack of discrimination information.Experimental results show that the human cascade detection network proposed in this paper can significantly improve the performance of occlusion pedestrian detection.